Speaker
Description
The environment of infrasound stations is characterized by mesoscale wind speed and temperature fluctuations that affect the temporal variability of the Atmospheric Boundary Layer (ABL). While the statistical characteristics of turbulence are poorly constrained, modeling such statistics appears to be crucial since each sensor of infrasound stations is subject to this local noise that may mask true signals and cause false detections. In this work, we propose to improve the station processing by characterizing the noise due to turbulence in the ABL using neural networks. Assuming that the turbulence is governed by a parametric nonlinear dynamical system which involves known dimensionless numbers, a neural network architecture is proposed to infer the turbulent noise in the data. For this task, we design a custom deep autoencoder network to obtain a coordinate transformation into a reduced space where the dynamics of the ABL can be sparsely represented. The resulting modeling framework combines the strengths of deep neural networks for flexible representation and sparse identification of nonlinear dynamics for parsimonious models. The performance of our approach is assessed using real-world signals recorded at several infrasound stations of the International Monitoring System, over days and nights, and for different seasons.
Promotional text
We introduce a new strategy to reduce the wind noise in the recorded signals of the IMS stations. This strategy is based on using machine learning to extract turbulence noise from data than can be translated into knowledge about the underlying fluid mechanics.
